SWE-bench measures an AI's ability to solve real GitHub issues — the kind of work engineers actually do. In early 2023, the best AI systems scored 1.96%. By mid-2026, that number crossed 72%. That is not incremental improvement. That is an inflection.
This doesn't mean AI is going to log into your Jira and ship features autonomously tomorrow. But it does mean the trajectory is not slowing down, and anyone in software who hasn't internalized what this curve implies is making a planning mistake. The AGI timeline data is what makes this urgent rather than theoretical.
Entry-level engineering has always been about volume tasks: fixing bugs with known patterns, writing boilerplate, implementing specs someone else designed, adding tests, doing code review on familiar codebases. These are precisely the tasks AI systems now perform reliably.
Hiring data has already shifted. Engineering headcount at mature tech companies started declining in 2024–25, and the percentage of that decline attributed to AI-assisted productivity (rather than macro conditions) has grown each quarter. Junior-to-senior ratios are thinning. This is not a 2030 problem.
The mid-level engineer role has always been a mixture: some execution (which is now highly automatable), some design judgment, some stakeholder communication. The execution portion — which is substantial — is already being compressed by AI tools. The fastest-growing use case is mid-level engineers using AI to do 3–4x the output, which means teams that used to need 4 people need fewer.
The risk here isn't elimination — it's that the role bifurcates. Either you're the person directing the AI, or you're increasingly redundant. There's less room in the middle.
Systems design, architecture decisions, debugging complex distributed systems, translating business requirements into technical constraints — these require contextual judgment that current AI systems handle poorly. They also require building trust with stakeholders over time, which doesn't compress the same way.
Senior engineers are not safe from AI — they're safe from the current wave. The 2031 timeline matters here. If reasoning AI arrives at the scale most forecasters now expect, the judgment and contextual complexity that makes senior engineering hard also becomes more tractable for AI. It's a longer runway, not an escape hatch. For a cross-profession view of who's most exposed, see the full profession breakdown.
howcloseisagi.com is read by engineers and founders who take the timeline seriously. One sponsor slot per post — no ad networks, no noise.
Get in touch →The engineers who will be most valuable in 2028–31 aren't the ones who avoided AI tools. They're the ones who understand AI systems deeply enough to direct, debug, and architect around them. The work shifts from writing code to designing systems that include AI components — which is harder, not easier, and requires more understanding of the underlying models, not less.
Practically, this means:
The SWE-bench number is tracking almost linearly with the broader capability curve. The same models that scored 72% on coding benchmarks have also crossed PhD-level performance on science reasoning tests, and near-human performance on professional bar exams and medical licensing. These aren't coincidences — they're the same underlying capability improvement reaching different domains.
The median AGI forecast is 2031. What that means for software engineering specifically: the 2027–28 inflection compresses junior work, the 2029–30 period likely automates mid-level execution substantially, and 2031+ is when the architecture and systems design work that currently feels safe starts to become contested terrain. It's not a cliff — it's a slope that's been getting steeper every year.